Multi-view Active Learning Based On Double Branch Network

Image classification,as one of the most popular topics in computer vision,plays a significant role in our life no matter in daily activities,medical diagnosis,or military activities.The popularity of the Internet and the update of electronic equipment make the image data grow explosively in both quantity and variety.These changes also provide higher research value and application value for image classification technology.With the arrival of big data and the development of artificial intelligence,the accuracy of recognition and classification based on deep learning has greatly improved.Although deep learning has subverted some traditional perceptions of computer version,this technology still has some shortcomings in practical applications.It requires a large amount of labeled data to train a well-performing model.On the one hand,the image itself is susceptible to factors such as lighting and occlusion which makes the validity of the image data different.On the other hand,image data usually doesn't have any label information,especially for a specific problem such as medical image.Therefore,it takes a lot of manpower,material resources and time to build a large and standardized dataset with annotations.How to effectively use the existing labeled data to label the unlabeled data and expand the labeled dataset efficiently is what we are concerned about.Another goal of this thesis is how to use part of the dataset to obtain the effect of entire dataset training.Regarding the above problems,we propose a multi-view active learning method based on double branch network.The main research of this thesis is as follows:1)We design a deep model which consists of two Bayesian convolutional neural network with different structures.We improve the performance of the model by combining the different characteristics of the two substructures to obtain more feature expressions.2)We propose a multi-view sampling strategy to select unlabeled data based on the double branch model.We combine the multi-level feature information of the deep model to select ‘valuable' samples to improve the performance of our model.The expansion of the training dataset and the training of our model are performed in an iterative manner.Experiments are conducted on two image datasets,Fashion-MNIST and Cifar-10.The results show that our proposed method achieved good performance.Our method can use existing labeled data to select unlabeled data for labeling,and effectively expand the training dataset.It improves the labeling situation caused by the lack of labeled samples in real scenes and obtains a higher classification accuracy.